Karpathy Uses Claude to Build 3D Tolkien Scene for $10

Why this matters Ten dollars and two hours to turn a paragraph of prose into a navigable 3D scene is a small number with big implications for how we judge model progress. Karpathy's benchmarks tend to stick, first the pelican-on-a-bicycle test, now Middle-earth, because they're cheap, visual, and hard to fake. That matters for developers and founders building on Claude or similar models: code generation is no longer just about passing unit tests, it's about whether a model can plan a scene, place objects, and animate them with only screenshot feedback, no real vision loop.

The mistakes Karpathy noted, errors introduced because Opus couldn't actually watch its own render, are the interesting part. They point to a concrete gap between token generation and spatial reasoning that researchers will need to close before "vibe coding" 3D worlds becomes reliable rather than a neat demo. Worth watching whether Anthropic or competitors respond with better self-checking tools, and whether Karpathy's informal tests get adopted as something closer to a standard.